Sebastian Raschka's from-scratch video course on building reasoning models

The course runs to eight lecture videos, over 11 hours in total, and the coding is done in Python and PyTorch. It begins from a small pretrained checkpoint, so it skips full pretraining. Topics include evaluation on MATH-500, inference-time scaling through sampling, self-consistency and self-refinement, reinforcement learning with verifiable rewards using a simplified GRPO loop, and distillation from a DeepSeek R1 teacher model. The videos sit alongside his Manning book, which can be found at Manning and Amazon, and they follow his earlier Coding LLMs from the Ground Up material. He also offers a shorter 90-minute LinkedIn Learning course covering the same concepts at a higher level. Commenters on Hacker News were mostly positive, though a smaller group questioned how useful from-scratch tutorials are in practice. For a team building an AI company, the course is a low-cost reference for how reasoning gains can come from extra compute at inference and from reinforcement learning on a small open model. Cost: Free. Founders can use it to set up answer verifiers and scoring pipelines before committing GPU budget to post-training.